Most enterprise AI budgets are being spent in the wrong places, and the deployment numbers are starting to prove it.
The artificial intelligence apps market is not monolithic. It’s fracturing into categories with wildly different adoption curves, different ROI profiles, and different risk postures. If you’re an IT director trying to rationalize your AI portfolio heading into next budget cycle, the category matters more than the vendor. Picking the right category of AI investment is the decision that compounds. Picking the wrong one is how you end up with a governance mess and nothing to show leadership.
Here’s what the adoption data and deployment patterns actually show.
The Categories Growing in Enterprise
Document intelligence is the clearest enterprise winner right now. Contract review, policy extraction, regulatory filing analysis — these are high-volume, high-stakes workflows where AI demonstrably reduces time and error rates. The ROI is calculable because the baseline cost is calculable. Legal teams, compliance functions, and procurement departments are signing off on these tools because the alternative is headcount they can’t hire fast enough.
Code generation and developer tooling is the second growth category, but with a caveat. Adoption is real. Productivity gains are documented in enough internal studies now that the skepticism has largely died. The caveat: the tools that are winning inside enterprises are not the consumer-facing ones. They’re the versions that can be configured to avoid training on proprietary code, constrained to internal repositories, and audited. The category is growing; the specific products winning are the ones that cleared security review.
Customer-facing AI agents are growing in deployment volume, particularly in financial services and telecommunications, where deflecting tier-one support volume has an immediate cost impact. The category is real. The implementation failure rate is also real, and most of the failures trace back to the same root cause: the AI had access to systems it shouldn’t have, or it couldn’t be audited when something went wrong.
The Categories Stalling — and Why
General-purpose AI assistants deployed enterprise-wide are underperforming expectations almost universally. The pitch was productivity at scale. The reality is inconsistent adoption, shadow IT proliferation when the sanctioned tool disappoints, and mounting questions about where the data goes. These artificial intelligence apps aren’t technically failing — they’re organizationally failing because they were bought without a control architecture underneath them.
Vertical AI applications built on third-party models are hitting a wall in regulated industries. Healthcare, finance, insurance, and government buyers wanted purpose-built AI for their domain. They got it — and then legal came back with questions about model training data, data residency, and audit trails that the vendors couldn’t answer cleanly. Deals are stalling at procurement, not at technical evaluation.
The pattern across stalling categories is consistent. The AI capability itself is often credible. What’s missing is the infrastructure layer that makes it operable in an enterprise context: access controls, data isolation, logging, and the ability to demonstrate compliance to an auditor who doesn’t care about your demo.
The Category with the Clearest ROI: AI Control Infrastructure
There’s a category of artificial intelligence apps that the analyst firms are starting to name but that enterprise buyers have been building around for two years already: AI control infrastructure. This is the layer that sits between your organization and whatever AI models or applications you’re running — enforcing policy, managing access, isolating data, and producing the audit records that make everything else defensible.
The ROI case for this category is not about AI capability. It’s about risk-adjusted deployment velocity. Organizations with a control infrastructure in place are deploying more AI applications, faster, because each new deployment doesn’t require starting a governance conversation from scratch. The infrastructure answers the governance questions by design.
This is the category Peridot was built for. The thesis is that enterprises don’t have an AI capability problem — the models are good enough for dozens of high-value use cases right now. They have a control problem. They can’t run AI inside their own infrastructure with the access management, data boundaries, and audit capability that regulated industries require. Peridot is the control layer that makes deployment defensible without slowing it down.
The enterprises that have made this investment are not running less AI. They’re running more of it, because every application they deploy is running inside a framework their security and compliance teams already approved.
What This Means for Your AI Portfolio
The practical implication for IT directors is this: your AI portfolio decisions should be evaluated not just on capability but on deployability. An AI application that can do 80% of what the best tool does, but that runs inside your infrastructure with full audit capability, is worth more in a regulated environment than the best-in-class tool that your legal team won’t clear.
This doesn’t mean accepting inferior AI. The gap between frontier models and deployable enterprise models has narrowed significantly in the past eighteen months. It means being honest that the bottleneck in most enterprise AI programs is not the model — it’s the surrounding infrastructure that makes the model usable at scale without creating new liability.
The artificial intelligence apps categories that are winning share one characteristic: they are either inherently controllable, or they are being deployed on top of control infrastructure. The ones stalling are the ones where the capability arrived before the governance did. That sequence — capability first, governance later — is expensive to reverse.
Peridot’s position is straightforward: enterprises that build control infrastructure first will deploy more AI, faster, with less regulatory exposure than those trying to retrofit governance onto applications already in production. The category isn’t a cost center. It’s the infrastructure that makes every other AI investment defensible, and that makes it the highest-leverage enterprise AI spend available right now.
The winners in enterprise AI adoption over the next three years won’t be the organizations that moved fastest. They’ll be the ones that built the infrastructure to move decisively — and kept moving when others had to stop and explain themselves to a regulator.